@article{ChongTanNg2007,
author = {Chong, Hui-Qing and Tan, Ah-Hwee and Ng, Gee-Wah},
affiliation = {Nanyang Technological University School of Chemical and Biomedical Engineering Nanyang Avenue Singapore 639798 Singapore},
title = {Integrated cognitive architectures: a survey},
journal = {Artificial Intelligence Review},
publisher = {Springer Netherlands},
issn = {0269-2821},
keyword = {Computer Science},
pages = {103-130},
volume = {28},
issue = {2},
url = {http://dx.doi.org/10.1007/s10462-009-9094-9},
note = {10.1007/s10462-009-9094-9},
year = {2007}
}
The original paper is available online: http://www.springerlink.com/content/n30n8n46468210q7/
All 6 have been applied to a variety of cognitive tasks [p123]
“is one of the first cognitive architectures proposed” [p106]
Based on classical AI; it is learning and experience driven with a focus on problem solving. [p106]
SOAR is used for the understanding and incorporation of intelligent behaviour mechanisms in classical AI. [p106]
A flow diagram of the SOAR system is provided:
+---------------------------------------------+
| Long Term Memories |
+------------+ |---------------------------------------------|
| Learning | |+------------+ +-----------+ +------------+|
| Mechanism |<-->|| Procedural | | Semantic | | Episodic ||
+------------+ || Memory | | Memory | | Memory ||
|+--------+---+ +-----+-----+ +--+---------+|
+------------+ | | | | |
| Decision | | v v v |
| Procedure |<-->| +-------------------------------+ |
+------------+ | | Working Memory | |
| +-----------------------+-------+ |
+--------------^---------------|--------------+
| |
| v
+------+-----+ +--------+
| Perception | | Action |
+------------+ ++-------+
^ |
| v
+---+---------+
| Environment |
+-------------+
The external environment state is made available through a perception module and can be influenced by implemented actions. [p106]
Long term memory stores procedural, semantic and episodic knowledge. [p106]
Working memory stores knowledge of goals, perceptions, hierarchy of states and operators relevant to current context. [p106]
Learning occurs when impasses arise; impasses are classified as no-change, tie, conflict, and rejection. [p107]
Learning occurs using chunking, reinforcement learning, episodic memory, and semantic memory techniques. [p107]
Used in problem solving task games, and by the US military for modeling, simulation and control. [p123]
A system that uses empirical cognitive psychology data and brain imaging to model human cognition [p107]
Step by step understanding and prediction tool for human cognition systems. [p107]
A flow diagram of the ACT-R system is provided:
+--------------------+ +------------------------+
| Intentional Module | | Declarative Module |
| (not identified) | | (temporal/hippocampus) |
+--------+^----------+ +---------+^-------------+
|| ||
+------v+-------+ +------v+----------+
| Goal Buffer | | Retrieval Buffer |
| (DLPFC) | +-----------------------+ | (VLPFC) |
+------+--------+ | Productions | +-------+----------+
+------------>| (Basal Ganglia) |<---------+
+------------>+-----------------------+<---------+
| | Matching (Striatum) | |
| | Selection (Pallidum) | |
| | Execution (Thalamus) | |
| +-----------------------+ |
+------+--------+ +-----+---------+
| Visual Buffer | | Manual Buffer |
| (Parietal) | | (Motor) |
+------^+-------+ +----^+---------+
|| ||
+-------+v--------+ +------+v------------+
| Visual Module | | Manual Module |
| (Occipital etc) | | (Motor/Cerebellum) |
+----------------^+ ++-------------------+
| |
++--------------------------------v+
| Environment |
+----------------------------------+
“The external environment and knowledge stored in the memories work conjunctively to select actions for execution to satisfy the goal(s) of the agent.” [p108]
As seen in the above figure, there are four basic modules, visual, manual, declarative memory and goals these are coordinated through the central production system, to enable cognition. [p108]
While the system is highly parallel it is limited to serial communication with each module, and the production system is only aware of the information in the serial buffers. [p108]
Buffers are limited to one declarative unit (chunk) at a time. [p109]
Procedural memory stores production rules and supports learning through production compilation. [p109]
Framework for Tower of Hanoi, memory for text or lists of words, language comprehension, and communication. [p123]
Used for military aircraft control and brain activity prediction. [p123]
Used for HCI research [p124]
A flow diagram of the ICARUS system is provided:
+------------+
| Long Term | +--------------+
| Conceptual | | Long Term |
| Memory | | Skill Memory |
+------^-----+ +---+--------+-+
| | |
| | |
+-------+--------+ +------v-----+ +v----------+
| Categorisation | | Skill | | Means End |
+----+-----^-----+ +------> Retrieval | | Analysis |
| | | +------+-----+ +----+------+
| | | | |
+-----v-----+-------+ | | |
| Short Term | | +--v------------v--+
| Conceptual Memory +-----+ | Short Term Skill |
| (Belief Memory) | | Memory |
+-------------------+ +------------------+
| Perceptual Buffer | | Motor Buffer |
+--------^----------+ +--------+---------+
| |
| |
+-----+-------+ +--------v--------+
| Perception <--+ | Skill Execution |
+-------------+ | +--+--------------+
| +------+
| |
++------------v-+
| Environment |
+---------------+
Rooted in physical and embodied agents, integrates perception and action with cognition. [p109]The 4 main modules are [p110-111]
“ICARUS has been applied to many cognitive tasks, including the Tower of Hanoi, multi-column subtraction, and peg solitaire. Other key domains, which have been studied to date, include in-city driving and pole balancing.” [p125]
A flow diagram of BDI is included:
+---------------------------+
| Database |
|---------------------------|
|+--------------++---------+|
|| Plan Library || Beliefs ||
|+-----+--------++---+-----+|
+------|-------------|------+
Plan as| +--+
Recipe | | Instansiated Plans
+v----------v---+ +------------+
+---------+ | <------> Intentions |
| Desires +-------> Interpreter | +------------+
+---------+ +----> | Selected Intentions
| ++------------+-+
Events| | Internal |
| | Actions |
+------+-----v+ |
| Event Queue | |
+--------^----+ +--v---------------+
| | External Actions |
+-------+----+ +--+---------------+
| Perception | |
+----^-------+ |
| |
| |
| +------------v--+
+------+ Environment |
+---------------+
Based on intentional systems and human practical reasoning theories. [p111]“The body of a plan comprises of possible courses of actions and procedures to achieve a goal” [p111]
The term desires refers to non-conflicted goals. [p112]
Intentions are a set of action used in a desire attempt. [p112]
The system evaluates context each cycle, and formulates new desires and plans to complete them. [p112]
Means-end reasoning is used in the context of the current intention, to reduce reasoning time. [p112]
Learning is not classically integrated, but others have presented ways to incorporate learning. [p113]
Originally designed for the reaction control system (RCS) in space shuttles. [p125]
Factory process control and business process management. [p125]
Performing cognitive tasks, such as the Tower of Hanoi. [p125]
Embodied conversational agents able to make small talk conversation and provide information. [p125]
Derived from behaviour-based robotics. [p113]
Decomposes problems based on behaviours exhibited while solving those tasks. [p113]
A hierarchy of competency layers with global access to context and allows parallel execution of behaviours. [p113]
A flow diagram of the subsumption architecture is included:
+ +
+ +---------+ +
+--->| Level 3 +------+
| +---------+ |
| |
| +---------+ |
+--->| Level 2 +------v---+
| +---------+ |
| |
| +----------+ |
+--->| Level 1 +---------v---+
| +----------+ |
| |
| +----------+ |
+--->| Level 0 +-------------v---+
| +----------+ |
| |
| |
| |
| |
+---------+ +-----v-----+
| Sensors | | Actuators |
+-----^---+ +---+-------+
| |
+-+--------------------------v-+
| Environment |
+------------------------------+
Each layer attempts to reach a goal, while being subsumed by higher level layers. [113]Some examples include Robots with basic navigation abilities, market simulation in games, reactive musical accompanist, co-ordinated soccer playing in RoboCup. [p125]
A hybrid architecture that uses both implicit (procedural, neural net learned) and explicit memories for reasoning and learning. [p114]
“able to react in a dynamically changing environment without any pre-existing knowledge installed into the architecture.” [p114]
A flow diagram of the CLARION is included:
Top Level
+---------------------------------------+
|+--------------+ +--------------+|
|| Explicit | | Explicit ||
+-->|Representation| |Representation|<--+
| +--^--^------+-----------------------^--+ |
| | | | | | | | | |
| | | | +--------------------+ | | |
| | | | | | | | | |
| | | +-------------------+ | | | |
| +--v----------------------+-------v--v--+ |
| || Implicit | | Implicit || |
+-->|Representation| |Representation|<--+
| |+--------------+ +--------------+| |
| +---------------------------------------+ |
| Bottom Level |
| |
| |
| +------------+ +---------+ |
+---+ Perception |<--+ +----+ Actions +----+
+------------+ | | +---------+
| |
| v
+----+----------+
| Environment |
+---------------+
Note that the dual Explicit/Implicit sections refer to the action centered (ACS) and the non-action centered subsystem (NACS) to account for variability in representation of explicit and implicit knowledge. [p115]
There are two levels, for the implicit and explicit knowledge and associated mechanisms; both layers are referenced during reasoning. [p115]
There are two subsystem not included in the diagram that motivate and control the system. [p115]
Rule based and similarity based reasoning are both used. [p116]
Learning can occur through procedural skills reinforcement or through explicit knowledge manipulation [p116]
“CLARION has been used in both the simulation of navigation and cognitive tasks.” [p125]
“The cognitive tasks using CLARION include serial reaction tasks, artificial grammar learning tasks, process control tasks, alphabetical arithmetic tasks, and the Tower of Hanoi.” [p125]
With the exception of the subsumption architecture, problem solving, reasoning/inference, and learning are essential components of most cognitive systems. [p126]
Many architectures use working memory in order to provide a manageable problem workspace incorporating the environment and relevant long term memory knowledge. [p126]
Most of the architectures use rules to dictate the actions that will be performed by the system. [p126]
Most of the architectures identify procedural (situational actions) and declarative (facts/inference rules] knowledge independently. [p126]
CLARION additionally includes episodic memory, could have more capacity to emulate human cognition. [p127]